Capture once.
Recall when it matters.
Recall Racoon gives your AI systems persistent memory. Information enters from multiple sources. Rumi Rummages through it. Your Stash remembers. Your AI recalls.
Built for production
AI memory
Every feature exists because we hit the same walls building agentic systems. No theoretical features—only what ships.
Deterministic Control Plane
Separate reasoning from memory policy. Safety rules, personality constraints, and authoritative facts are enforced at the control plane, never left to probabilistic drift.
Hybrid Semantic Search
Vector + BM25 + RRF reranking with local ONNX embeddings. Works fully offline. No API keys, no per-token fees, no data leaves your infrastructure.
Claim Lifecycle Governance
Every memory is a lifecycle-managed claim: candidate → confirmed → stale → superseded → conflicted → archived. Full provenance, citation tracking, and human-in-the-loop review.
MCP-Native Integration
Drop into Claude Desktop, Cursor, VS Code, or any MCP client. Thirteen tools exposed over stdio, SSE, or HTTP. Zero config required.
Local-First by Default
SQLite + sqlite-vec for fully offline operation. Optional Qdrant, Postgres, or Cloudflare Workers backends when you need scale. Your data, your rules.
Secret Redaction & ACL
Auto-redact JWTs, AWS keys, SSH keys, and Bearer tokens at ingest. Memory-level access control with composite IDs and group-scoped search spaces.
Persistent Ingestion
Accept documents, API payloads, agent outputs, and chat transcripts. Automatic chunking, embedding, and enrichment at write time.
Namespaces & Tags
Organize memories by project, team, environment, or custom taxonomies. Scoped search prevents cross-contamination.
Short & Long Term Memory
Separate working memory (session-scoped) from persistent Stash. Automatic promotion and decay policies.
Plug into your
existing stack
Recall Racoon speaks the protocols your agents already use. No SDK lock-in, no vendor-specific APIs.
MCP Protocol
Native Model Context Protocol implementation. 13 tools: ingest, search, graph, steward, and admin operations.
REST API
Full HTTP API for custom integrations. OpenAPI spec included. Webhook support for async operations.
Document Import
PDF, Markdown, HTML, JSON, and plain text ingestion. Automatic metadata extraction and structure preservation.
Agent Frameworks
Pre-built adapters for LangChain, LlamaIndex, AutoGen, CrewAI, and custom agent loops.
How it works
under the hood
1Ingestion & Enrichment
Documents and data enter via MCP tools, REST API, or direct SDK. Rumi immediately applies format validation, secret redaction, and metadata extraction. Each piece becomes a candidate claim with full provenance.
2Rummage Processing
Rumi runs multi-probe validation: format, citation, semantic, and contradiction probes. Claims pass through a 5-tier resolution: auto-confirm, flag for review, supersede, archive, or reject. All decisions are auditable.
3Stash Persistence
Confirmed claims write to the Data Plane (SQLite/Postgres + vector index). Append-only with snapshots. Namespaces provide isolation. Embeddings stored locally via ONNX—zero external dependencies.
4Recall & Reuse
Hybrid search (vector + BM25) with RRF reranking. Freshness scoring, confidence thresholds, and namespace scoping. Results include citations and graph context. Sub-50ms latency typical.